2.1.1 Experimental object and equipment

A commercial cylindrical LFP battery (WTT32700, Shenzhen Nuoxiang Electronics Co., Ltd., Shenzhen, China) with a nominal capacity of 6 Ah and a nominal voltage of 3.2 V was selected for this study. It is suitable for high-rate discharge applications. The main parameters of the battery are summarized in Table 1.

Charge-discharge tests were conducted using a battery test system (CT-4008-5V20A-A, Neware Technology Limited, Shenzhen, China). The control accuracies for voltage and current were ±0.02% and ±0.05%, respectively, and data were recorded at a frequency of 10 Hz. The ambient temperature was maintained using a low- and high-temperature thermostatic bath (DC-2006, Shanghai Zhulan Instrument Technology Limited, Shanghai, China), with a temperature fluctuation of ± 0.1℃.

2.1.2 Internal pressure sensing and acquisition system

Internal pressure measurement constitutes a core innovation of this study. A SUP-P300-B (Hangzhou Meacon Automation Technology Co., Ltd., Hangzhou, China)pressure transmitter was employed for this purpose, and its main parameters are summarized in Table 2.

The sensor is isolated from the measured medium by a polytetrafluoroethylene (PTFE) coating and a stainless-steel corrugated diaphragm, enabling long-term operation in harsh environments, such as acidic, alkaline, oily, and gaseous media. When pressure is applied to the sensor, it outputs a voltage signal proportional to the pressure. After amplification and calibration, the output voltage is linearly related to the pressure value.

In the experiment, the sensor was connected to the battery via a custom sealed fixture with sealing to directly measure the internal gas-phase pressure of the battery. The sensor had a measurement range of 0–200 kPa, an accuracy class of 0.5, and a 4–20 mA current output. The pressure data were converted and synchronously recorded using a paperless recorder (MIK-R200T, Hangzhou Meacon Automation Technology Co., Ltd., Hangzhou, China).

The internal pressure curve obtained in this study exhibits a more pronounced cumulative trend than those reported in some existing studies. This discrepancy arises primarily from differences in cell type and measurement methodology. Most previous investigations measure the external applied pressure on pouch cells using external clamping fixtures. In the present work, experiments were conducted on cylindrical 32700 cells, with the sensor directly connected to the internal gas phase via a custom sealing fixture. The rigid steel casing of the cylindrical cell provides a nearly constant volume, and the internal pressure is governed by gas accumulation from side reactions and the volume changes of active materials during lithium insertion and extraction. This closed and constrained space produces a more noticeable pressure build-up effect than that observed in pouch cells, where external pressure measurements are heavily influenced by mechanical deformation of the soft casing. Therefore, the curve morphology in this study reflects the intrinsic gas-phase pressure response of cylindrical LFP cells, which is qualitatively consistent with the underlying electrochemical-mechanical coupling mechanism.

2.2.1 Experimental data acquisition and dynamic operating condition design

To extract informative signals from the internal pressure data, a resting temperature-pressure calibration experiment was first conducted to characterize the influence of temperature on internal pressure. Subsequently, multi-temperature and multi-rate pulse experiments were performed to capture internal pressure responses under dynamic operating conditions. The measured internal pressure was then combined with temperature, voltage, and current as input features for the model.

To verify that the pressure variations primarily originate from electrochemical reactions rather than elastic deformation of the cell casing, a mechanical analysis based on thin-walled cylinder theory was conducted (see Supporting Information). The results indicate that the volume change of the casing is negligible within the tested pressure range (< 0.1‰), confirming that the internal pressure signal directly reflects variations in the amount of internal gas. Therefore, over the experimental pressure range, the volume change induced by elastic deformation of the casing is negligible compared with the internal gas volume, and the gas volume can be approximately treated as constant.

Under resting conditions without electrochemical reactions, the total amount of internal gas and the casing volume can be assumed constant. The ideal gas equation P = nRT/V_gas can be used to determine the relationship between pressure (P) and temperature (T), where R represents the molar gas constant, V_gas represents the volume, and n represents the amount of substance. During battery operation, however, n may vary dynamically due to electrochemical processes such as lithium insertion/extraction and gas generation from side reactions, making it a key variable related to SOC. In contrast, changes in T induce thermodynamic pressure fluctuations and act as a non-electrochemical disturbance. If raw pressure is directly used as a model input, the model may learn both temperature-induced disturbance and SOC-related information, which can degrade generalization under wide-temperature operating conditions. Therefore, it is necessary to identify the pressure component associated with electrochemical processes while accounting for the influence of temperature.

The experimental data consist of two parts. The first part is the resting temperature-pressure calibration experiment, which was used to analyze the influence of temperature on battery internal pressure. Because the target temperature of electric-vehicle thermal management systems is typically set within 25℃–35℃, a heating-rest experiment was conducted for the battery at SOC = 100% and without external charge-discharge excitation over a temperature range of 25℃ to 36℃. After the battery reached thermal equilibrium, the corresponding temperature and internal pressure data were recorded. This experiment was used to characterize the basic variation of internal pressure with temperature under resting conditions.

The second part is the multi-temperature and multi-rate pulse experiment, which was used to obtain the voltage, current, temperature, and internal pressure responses during dynamic charge-discharge operation. At three constant temperatures of 25℃, 35℃, and 45℃, the battery was first fully equilibrated to ensure that the battery temperature was stable and consistent with the ambient temperature. It was then discharged at a constant current of 0.2C to the cut-off voltage of 2.0 V and rested for 1 h, and this state was taken as the initial low-SOC state. The battery was then charged at 0.2C to different SOC reference points (0%, 10%, ..., 90%). At each SOC reference point, 180 s constant-current charge pulses and 180 s discharge pulses at 0.1C, 0.2C, 0.5C, and 0.8C were sequentially applied, with a 3 min rest interval between pulses, to obtain dynamic responses under different current excitations. The multi-temperature pulse test and data recording procedure is shown in Figure 1A. During the experiments, current (I), terminal voltage (V), temperature (T), and measured internal pressure (P) were synchronously collected, and SOC during the dynamic process was corrected in real time using Coulomb counting. Thus, a raw time-series dataset covering a wide SOC range, multiple temperatures, and multiple C-rate conditions was constructed.

Figure 1 Overview of the experimental setup and model framework. (A) Data acquisition procedure; (B) temperature-pressure fitting curve; (C) overall architecture of the proposed model

Three identical WTT32700 batteries were tested to verify data consistency. Approximately 120,000 time-step samples were collected at each temperature. The entire dataset was divided chronologically into training (70%), validation (15%), and testing (15%) sets. This chronological split ensures that the test data are temporally independent of the training process and prevents data leakage that could arise from random shuffling of adjacent samples. Consequently, the evaluation reflects the model's true generalization capability to unseen conditions, rather than mere interpolation of the training data.

In addition, to rigorously verify the model's wide temperature domain generalization ability, two complementary strategies were designed to evaluate the generalization ability. Domain adaptation validation: a base model was first trained on the full 25℃ dataset and then fine-tuned using a small portion of initial samples from the target temperature (15% of the total target data). The test was performed on the remaining 85% of the target temperature sequences. This setup simulates a practical scenario in which a model is rapidly adapted to a new condition with limited target-domain samples. Cross-temperature boundary test: as a limit exploration of generalization performance, a model trained exclusively on 25℃ data was directly tested on the full 35℃ and 45℃ datasets (zero-shot testing). Although the accuracy decreases due to the distribution shift, this test reveals the true impact of temperature domain shift and provides a benchmark for future domain adaptation research.

Together, these two strategies assess the model's adaptability across temperatures: the domain adaptation validation demonstrates practical transfer performance, while the boundary test reveals the limitations of purely data-driven methods in zero-shot generalization.

2.2.2 Internal pressure signal analysis and input feature construction

The internal pressure of the battery is influenced by both temperature variations and electrochemical reactions. Under resting conditions, the experimental data reveal a pronounced nonlinear relationship between internal pressure and temperature, as shown in Figure 1B. Based on the characteristics of the curve, an exponential function was adopted for empirical fitting to establish the pressure-temperature relationship:$P_{\text{s}\text{t}\text{a}\text{t}\text{i}\text{c}}=0.3569\times e^{T/10.63757}+20.22188$(1)where P_static denotes the theoretical static internal pressure; T is the battery temperature.

The fitted model yields a coefficient of determination r² = 0.99578, indicating that it accurately captures the pressure-temperature relationship over the experimental temperature range.

Under dynamic charge-discharge conditions, the internal pressure comprises not only a thermodynamic contribution induced by temperature but also the pressure response arising from electrochemical-mechanical processes, such as lithium insertion/extraction, volume changes of active materials, and possible gas generation from side reactions. Consequently, the internal pressure itself can reflect changes in the internal battery state and provide complementary information for SOC estimation beyond voltage signals. Given that temperature also affects internal pressure variations, measured P and T were simultaneously used as model inputs, enabling the LSTM network to learn the coupling among internal pressure, temperature, and SOC during temporal modeling.

Finally, I, V, T, and P served as input features, with the SOC reference value serving as the output label. All input features were normalized to [0, 1] using the min-max method to improve training stability and convergence speed.

2.3.1 Input and feature layer

At each time step t, the input to the model is a four-dimensional feature vector X_t = [I_t, V_t, T_t, P_t], representing the current, voltage, temperature, and internal pressure at time t, respectively.

2.3.2 LSTM temporal feature extraction layer

Two stacked LSTM layers were employed as the core temporal feature extractor. Through its gating mechanisms, the LSTM effectively captures long-term dependencies. The computation proceeds as Equations (2–7):$F_t=\sigma \left(x_t{W}_{\text{xf}}+h_{t-\text{1}}{W}_{\text{hf}}+b_{\text{f}}\right)$(2)$I_t=\sigma \left(x_t{W}_{\text{xi}}+h_{t-\text{1}}{W}_{\text{hi}}+b_{\text{i}}\right)$(3)$O_t=\sigma \left(x_t{W}_{\text{xo }}+h_{t-\text{1}}{W}_{\text{ho}}+b_{\text{o}}\right)$(4)$C_t^{\text{'}}=\tanh \left(x_t{W}_{\text{x}\text{c}}+h_{t-\text{1}}{W}_{\text{h}\text{c}}+b_{\text{c}}\right)$(5)$C_t=F_t\odot C_{t-\text{1}}+I_t\odot C_t^{\text{'}}$(6)$h_t=O_t\odot \text{t}\text{a}\text{n}\text{h}(C_t)$(7)where F_t, I_t, and O_t denote the outputs of the forget gate, input gate, and output gate, respectively; $C_t^{\text{'}}$ represents the candidate memory cell; C_t and C_t−1 are the current and previous memory cells; h_t and h_t−1 are the current and previous hidden states; σ is the sigmoid activation function; b denotes the bias term, and b_f, b_i, b_o, b_c are the bias vectors for the forget gate, input gate, output gate, and candidate memory cell, respectively; W denotes the weight matrix, and W_xf, W_xi, W_xo, W_xc represent the weight matrices for the forget gate, input gate, output gate, and candidate memory cell, respectively, and W_hf, W_hi, W_ho, and W_hc denote the weight matrices from the hidden state to the forget gate, input gate, output gate, and candidate state, respectively; x_t is the input at the current time step t; and ⊙ represents the Hadamard product.

2.3.3 Dual-attention mechanism

To enhance the model's capability to focus on key information and features, a dual-attention mechanism is employed. The two attention modules are sequentially stacked after the LSTM layers to jointly perform adaptive selection and aggregation of temporal features.

(i) Feature attention

In the SOC estimation task, the four input features: voltage, current, temperature, and internal pressure—contribute unequally across different SOC regions and operating conditions. For instance, in the voltage plateau region, terminal voltage exhibits low sensitivity to SOC variations, which increases the relative importance of current and internal pressure. A feature-attention mechanism was therefore introduced to enable the model to adaptively learn dynamic weights for different features at each time step.

Let H[h₁, h₂, ……, h_t]∈R^L×d denote the hidden-state sequence output by the LSTM layers, where L is the number of time steps and d is the hidden-state dimension. At each time step t, the hidden state h_t is mapped to the feature-attention space through a fully connected layer, as defined in Equation (8):$e_t=\text{tanh}\left(W_{\text{f}}{h}_t+b_{\text{f}}\right)$(8)where e_t is the feature-attention hidden representation at time step t, is a trainable weight matrix, d_f is the attention hidden dimension.

Subsequently, the feature-attention score is computed using another weight vector v, as shown in Equation (9):$s_{t\text{,}i}=v_{\text{f,}i}^L{e}_t$(9)where $v_{\text{f,}i}^L$ denotes the trainable weight vector for the i-th feature dimension.

To obtain normalized feature-attention weights, the softmax function is applied across different feature dimensions at each time step, as shown in Equation (10):${\alpha }_{t\text{,}i}=\frac{\text{exp}\left(s_{t\text{,}i}\right)}{{\displaystyle {\sum }_{j=1}^{d_{\text{f}}}\text{exp}\left(s_{t\text{,}j}\right)}}$(10)where α_{t, i} represents the attention weight of the i-th feature dimension at time step t, s_{t, j} represents the attention score of the j-th feature dimension at time step t, and s_{t, i} is the feature-attention score of the i-th feature dimension at time step t.

The weighted feature representation is then obtained via the Hadamard product, as given in Equation (11):${\tilde{h}}_t={\alpha }_t\odot e_t$(11)where ⊙ denotes the Hadamard product, α_t is the feature-attention weight vector at time step t. The output sequence of the feature-attention module is denoted by $\tilde{H}[{\tilde{h}}_1,{\tilde{h}}_2,\cdots ,{\tilde{h}}_t]$.

(ii) Temporal attention module

Battery SOC is a dynamically evolving state variable that depends not only on the current input but also on historical states. The temporal-attention mechanism is designed to automatically identify and attend to the historical time steps most critical for the current estimation.

The hidden-state sequence $\tilde{H}[{\tilde{h}}_1,{\tilde{h}}_2,\cdots ,{\tilde{h}}_t]$ weighted by the feature-attention module is fed into the temporal-attention layer. For each time step t, the temporal-attention score u_t is computed as Equation (12):$u_t=v_{\tau }^L\text{tanh}\left(W_{\tau }{\tilde{h}}_t+b_{\text{t}\text{e}\text{m}}\right)$(12)where $v_{\tau }^L$ and $W_{\tau }\in R^{d_{\text{f}}\times d_{\text{t}\text{e}\text{m}}}$ are the trainable weight vector and matrix, respectively; ${\tilde{h}}_t$ denotes the feature representation at time step t after feature-attention weighting; b_tem is the bias term, and d_tem is the temporal-attention hidden dimension.

The softmax function is then applied to normalize the scores across all time steps, yielding the temporal-attention weights β_t, as defined in Equation (13):${\beta }_t=\frac{\text{exp}\left(u_t\right)}{{\displaystyle {\sum }_{k=1}^L\text{exp}\left(u_k\right)}}$(13)where u_k denotes the score at the summation index k.

The weight β_t reflects the relative importance of the hidden state at the t-th time step for the current SOC estimation. Finally, the hidden states of all time steps are weighted and summed according to their temporal-attention weights to generate a comprehensive context vector c, as shown in Equation (14):$c={\displaystyle {\sum }_{t=1}^L{\beta }_t}{\tilde{h}}_t$(14)

The context vector c aggregates the most critical information across the entire sequence, preserving the dynamic semantics of multiple features while emphasizing the contributions of key time steps.

2.3.4 Output and model training

A fully connected layer with a sigmoid activation function maps the context vector produced by the dual-attention module to the estimated SOC. The model was trained with the mean squared error (MSE) loss and Adam optimizer.

To determine the model architecture and key hyperparameters, a grid search was conducted using the validation-set root mean square error (RMSE) as the evaluation criterion. Specifically, the number of LSTM layers was chosen from {1, 2, 3}, the number of hidden units from {32, 64, 128}, and the dropout rate from {0.1, 0.2, 0.3}.

The results indicate that a single LSTM layer is insufficient for capturing long-term dependencies under complex operating conditions, whereas three LSTM layers, despite increasing model capacity, also lead to longer training times and a higher risk of overfitting. In comparison, two LSTM layers achieve a favorable balance between expressiveness and training stability. With 64 hidden units, the model effectively extracts temporal information from multi-source features such as voltage, current, temperature, and internal pressure; increasing this to 128 yields only a marginal reduction in validation error while increasing model complexity. A dropout rate of 0.2 effectively mitigates overfitting and ensures stable convergence. Accordingly, the configuration consisting of two LSTM layers, 64 hidden units, and a dropout rate of 0.2 was adopted. The final hyperparameter settings are summarized in Table 3.

3.2.1 Estimation accuracy under individual temperatures

To further verify the advantage of the proposed LSTM-AM model over conventional static methods, the polynomial surface fitting model and the LSTM-AM model were evaluated on the same test set. As shown in Table 4, under the single-temperature static condition, the LSTM-AM model achieved a 24% reduction in RMSE and a 13% reduction in mean absolute error (MAE) compared with the polynomial model, and its r² value was closer to 1. These results indicate that even in a relatively simple static mapping scenario, the LSTM-AM model, owing to its stronger nonlinear representation capability, outperforms conventional fitting methods based on manually specified functional forms. More importantly, SOC estimation in practical applications is typically conducted under dynamic charge-discharge and varying-temperature conditions, whereas the polynomial model ignores historical temporal information and requires parameter refitting for each temperature, hindering its adaptability to complex dynamic environments. Temporal models such as LSTM are therefore necessary to learn the dynamic mapping between multi-source input features and SOC.

The performance of the LSTM-AM model was evaluated on test sets at 25℃, 35℃, and 45℃. Comparisons were also conducted with the transformer and CNN-LSTM models at 25℃. Figure 3A presents the estimated SOC against the true values over a full cycle for a standard LSTM at 25℃, while Figure 3B–D displays the LSTM-AM results at 25℃, 35℃, and 45℃, respectively. Figure 3E, F shows the corresponding results for the CNN-LSTM and transformer models at 25℃. The estimated SOC curve of the LSTM-AM closely follows the true curve, tracking SOC accurately even under challenging conditions, such as the voltage plateau region and large current transients, with no appreciable lag or drift.

Figure 3 Performance evaluation of models under different conditions. (A) LSTM at 25℃, (B) LSTM-AM at 25℃, (C) LSTM-AM at 35℃, (D) LSTM-AM at 45℃, (E) CNN-LSTM at 25℃, and (F) transformer at 25℃

The error metrics are summarized in Table 4. The RMSE of the LSTM-AM model remains below 1.5%, and its MAE is below 1% across all three temperatures. The model performs best at 35℃, because this temperature is close to the optimal operating point of the battery. At 25℃ and 45℃, the performance fluctuates slightly but remains excellent.

3.2.2 Effectiveness analysis of the attention mechanism

To validate the dual-attention mechanism, the proposed LSTM-AM was compared with an identical LSTM model without attention layers. As shown in Table 4, with the attention mechanism incorporated, the RMSE at 25℃ drops from 1.80% to 1.44% (a 20.0% reduction). At 35℃, it drops from 1.45% to 1.14% (a 21.4% reduction), and at 45℃, from 1.72% to 1.34% (a 22.1% reduction). This consistent improvement across temperatures demonstrates that adaptively weighting key features and historical states is universally beneficial, enhancing not only accuracy but also tracking stability in the voltage plateau region and during large current transients.

3.2.3 Error distribution and robustness analysis

To further evaluate the stability of the model estimates, error histograms and cumulative distribution curves were constructed for the 25℃, 35℃, and 45℃ test sets. Figure 4A–D presents the error histograms for the LSTM baseline without attention at 25℃ and for the LSTM-AM model at 25℃, 35℃, and 45℃, respectively. The errors approximately follow a normal distribution centered near zero, and more than 99% of the samples exhibit an absolute error below 3%. This indicates that the model is free of systematic bias and achieves high prediction consistency.

Figure 4 Error histograms of (A) the LSTM baseline model without attention at 25℃, the LSTM-AM model at (B) 25℃, (C) 35℃, and (D) 45℃, and (E) error distribution analysis

Figure 4E compares the error distributions before and after the incorporation of the attention mechanism. The LSTM-AM model exhibits a median error closer to zero, a narrower dispersion range, and fewer outliers. This confirms that the attention mechanism effectively improves both robustness and estimation accuracy by focusing on key information. The slightly broader error distribution observed at 45℃ can be attributed to the increased nonlinearity of the battery at high temperatures, where the relationships among SOC and the voltage, current, and pressure features become less stable due to intensified side reactions, larger pressure fluctuations, and a more ambiguous SOC mapping. Nevertheless, the errors remain within an acceptable range.

3.2.4 Generalization validation on a public dataset

To further verify the cross-domain generalization performance, the pre-trained LSTM-AM model was evaluated on the NASA lithium-ion battery public dataset and compared with three baselines: LSTM, CNN-LSTM, and transformer. The NASA dataset contains 504 complete discharge cycles from three 18650 LFP batteries (B0005, B0006, and B0007) at room temperature. The recorded signals include voltage, current, temperature, and average internal pressure (one constant value per cycle). Since the NASA dataset does not provide a temperature-pressure calibration curve, the raw average pressure value was used as an input feature. All models were first pre-trained on the 25℃ dataset, then fine-tuned using only the first 10% of the NASA samples and tested on the remaining 90%.

Figure 5 compares the SOC estimation results of the four models on 200 randomly selected test points from the NASA dataset. Figure 5A–D corresponds to the fine-tuned LSTM-AM, LSTM, CNN-LSTM, and transformer, respectively. Table 5 summarizes the RMSE, MAE, and r² for each model.

Figure 5 SOC estimation results of the four models on the NASA dataset. (A) LSTM-AM, (B) LSTM, (C) CNN-LSTM, and (D) transformer

The SOC estimation errors of all models on the NASA dataset are significantly higher than those on our own data. This discrepancy is primarily attributed to differences in battery chemistry and capacity, the transition from pulse discharge to continuous discharge profiles, and the limited availability of pressure information. A comparison among the models reveals that the LSTM-AM achieves an RMSE of 6.25%, outperforming the standard LSTM's 7.04% with a 11.2% relative reduction. This confirms that the dual-attention mechanism can effectively identify key features and time steps in cross-domain scenarios. The RMSE values of CNN-LSTM and transformer are slightly lower than that of LSTM-AM; the three models exhibit comparable MAE and r². This suggests that when feature information is incomplete, the choice of deep temporal architecture has less impact on accuracy than feature completeness. Nevertheless, LSTM-AM has a clear interpretability advantage over CNN-LSTM and transformer because its attention weights can measure the contribution of each input feature across many SOC regions, offering clear explanations of model behavior.

It is worth noting that the validation error on the NASA dataset is substantially higher than that obtained on the in-house dataset (RMSE 1.44 %). This increase is primarily attributed to the loss of detailed internal pressure information and differences in current signal characteristics. Subsequent ablation experiments have demonstrated that removing internal pressure P increases RMSE by 0.72 percentage points, while removing the current I leads to catastrophic performance degradation. The NASA dataset provides only a single average pressure value per cycle, and the current signal distribution differs due to the different battery capacities; this accuracy drop is therefore expected. Even so, the LSTM-AM maintains an r² above 0.93 using only voltage, temperature, and a coarse average pressure, indicating that the model can automatically extract robust SOC-related temporal features from multi-source signals rather than relying solely on internal pressure. This also suggests that in practical deployment, when accurate temperature-compensated internal pressure signals are unavailable, the model can be quickly adapted to a new battery system and operating conditions by fine-tuning with a small number of target-domain samples.

3.3.1 Feature contribution ablation experiment

Ablation experiments were performed on the 25℃ dataset to quantify the contribution of each input feature. Using the full four-feature LSTM-AM model as a baseline, individual features were removed one at a time. The results were shown in Figure 6 and Table 6.

Figure 6 Results of the ablation experiments. (A) LSTM-AM without V, (B) LSTM-AM without I, (C) LSTM-AM without T, and (D) LSTM-AM without P

The experimental results demonstrate that electrical signals constitute the primary inputs for SOC prediction, and the removal of voltage or current leads to severe performance degradation. Temperature also plays a significant role; excluding the temperature feature increases the RMSE by 0.34 percentage points. Removing internal pressure increases the RMSE by 0.72 percentage points, corresponding to a relative increase of 50%. Although this contribution is smaller than that of electrical signals, the improvement provided by internal pressure is a meaningful, supplementary gain.

This indicates that even when accurate electrical and thermal features are available, the internal pressure signal P still contributes unique information that enhances estimation accuracy. In the voltage plateau region, internal pressure P supplies independent and complementary SOC-identification information that cannot be supplied by terminal voltage alone.

3.3.2 Cross-temperature generalization performance validation

The model's adaptability to unknown temperatures was systematically evaluated using the two strategies described in Section 2.2.1. The cross-temperature boundary test, a zero-shot limit evaluation, involved training solely on 25℃ data and testing directly on the complete 35℃ and 45℃ datasets. The results are summarized in Table 7.

The zero-shot RMSEs reache 7.88% at 35℃ and 17.95% at 45℃, representing a substantial degradation relative to models trained on the full datasets at those temperatures. This performance drop reveals the fundamental impact of temperature domain shift on purely data-driven models: the coupling relationships among internal pressure, temperature, and SOC undergo structural changes that hinder generalization from 25℃ samples alone. Nevertheless, these results establish a valuable benchmark, quantifying the actual influence of temperature shift and confirming the necessity of domain adaptation for cross-temperature deployment. In practice, a seasonal or regional calibration strategy could be employed to adapt the model using only a small number of samples. The domain adaptation validation yields a very different outcome. After pre-training on full 25℃ data and fine-tuning with the first 15% of target temperature data, the model performs excellently. The LSTM-AM achieves an RMSE of 1.55% at 35℃ and 1.97% at 45℃. These results are very close to those obtained by training on the full target temperature datasets. This confirms that the temporal features and attention patterns learned at 25℃ are highly transferable and that only a small amount of target-domain data is required to achieve high estimation precision.